A new report published Thursday by the AI safety research organization METR details an 11-week investigation into Anthropic’s development practices. According to the findings, Anthropic is employing AI systems to build more powerful AI models, and this recursive approach appears to be functioning effectively. The research indicates that the company’s models are improving approximately twice as quickly as they did a year ago, showing minimal signs of deceleration. The Verge reported on the release of the METR findings.
Recursive Improvement and AI Safety
The practice of artificial intelligence systems assisting in their own development has long been discussed within the tech industry as a potential milestone toward rapid capability acceleration. When models can reliably optimize or generate succeeding iterations of themselves, the timeline for capability gains can theoretically compress. METR’s investigation provides external documentation of this dynamic occurring inside a major AI lab, raising questions among researchers regarding the pace and predictability of future model generations.
Unknowns and Future Monitoring
Despite the insights provided by the 11-week investigation, substantial details remain absent from the public domain. The report and subsequent coverage did not specify which internal models were utilized for the automated development process, nor did they disclose the precise engineering architectures enabling the reported acceleration. Furthermore, it remains unknown how Anthropic plans to manage safety evaluations as the feedback loop between AI systems and their iterative improvements continues to speed up.
Observers will be monitoring future disclosures from independent safety organizations and labs to see how recursive development methods scale over the coming year.
Lead image: Brett Sayles / Pexels
